Doctoral Dissertation Research: Recovering the Polyvalent Genealogies of Machine Learning, 1948 - 2017

点击翻译按钮获取中文摘要

Machine learning techniques currently make "high-stakes" judgments in areas as diverse as criminal justice, credit risk, social welfare, hiring, and congressional redistricting. Such techniques make these decisions using patterns learned from historical social data. Emphasis on prediction rather than the circumstances of dataset creation have led to machine learning systems that preferentially target vulnerable populations for disparately adverse social judgments while making it more difficult for those subject to these decisions to protest unfair treatment. This study explores the limitations of such machine learning systems by tracing how technical and non-technical people, including funding agencies, have historically understood what machine learning systems could and should achieve. Particular care is given to the forms of "learning" valued by researchers during different moments in the 20th century, and to the emergence of theoretical concepts that were constrained and even defined by the capabilities of the available material devices. This work makes visible the efforts of women and men previously omitted in histories of artificial intelligence and machine learning, and develops a quantitative method to document how the innovations of a discipline are contingent upon interdisciplinary and transdisciplinary research networks. Finally, this project traces how the allocation of resources to particular research communities spurred scientific innovation in adjacent and seemingly unrelated academic research fields in the physical and social sciences. In this sense, the discipline of machine learning provides a useful case study for modeling the propagation of ideas across different subfields.Both qualitative and quantitative historical research is employed. First, nine university and government archives are perused to reconstruct the institutional organizations, interpersonal research networks, and material computing devices available to machine learning and artificial intelligence researchers. Second, an investigation is conducted using a novel combination of topic modeling and word embeddings on a corpus of millions of full-text Association of Computing Machinery articles from 1950 to 2017 to trace how discursive influence propagates across disparate sub-disciplines. Four research products are generated: (1) a technical machine learning publication detailing the novel method used to analyze the article corpus, (2) a history of science article tracing the early history of machine learning, (3) a general audience "think piece" discussing the policy and ethical implications of contemporary machine learning research, (4) and the public release of the project code and the subsequent statistics generated from the article corpus. Digital copies of salient archive records discovered during this research study will be made freely available via Columbia University's Digital Repository, insofar as this is possible given the copyright and access restrictions of holding institutions. Archive materials collected and computational study of the article corpus will be used in the co-PI's doctoral dissertation exploring how machine learning has been used to classify individuals, imagine communities, and legislate forms of social and political evidence.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.